vlmo/torchscale/component/xmoe/routing.py

# Copyright (c) 2022 Microsoft
# Licensed under The MIT License [see LICENSE for details]

# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
#
# This source code is licensed under the BSD license found in the
# LICENSE file in the root directory of this source tree.

# Implementation of Top2Gating described in https://arxiv.org/pdf/2006.16668.pdf
# Code is inspired by Top2GatingOnLogits from lingvo:
#   https://github.com/tensorflow/lingvo/blob/21b8106c5f1d30a196c98eedc441d4fd70833b11/lingvo/core/moe_layers.py#L477

# NOTE: This is a mirror of the code in
# https://github.com/facebookresearch/fairscale/tree/master/fairscale/nn/moe

import math
from typing import Callable, Dict, Optional, Tuple

import torch
import torch.nn.functional as F
from torch import Tensor

from .moe_layer import fused_cumsum_sub_one, has_tutel

# use a fixed temperature to compute balance loss
TEMPERATURE_FOR_L_UAX = 0.07

# maximum capacity of 1 expert as a fraction of number of tokens in the batch
# Note: setting this to 1.0 causes inference to significantly slow down
EVAL_CAPACITY_TOKEN_FRACTION = 0.25

# logging
SAMPLE_FRACTION = 0.2


def top1gating(
    logits: torch.Tensor,
    input_mask: Optional[torch.Tensor] = None,
    use_fp32=False,
    capacity_factor=1.0,
    eval_mode=False,
    moe_eval_capacity_token_fraction=EVAL_CAPACITY_TOKEN_FRACTION,
    use_xmoe=False,
    gate_obj=None,
) -> Tuple[Tensor, Tensor, Tensor, Dict]:
    """Implements Top2Gating on logits."""
    metadata = {}
    if use_fp32:
        orig_dtype = logits.dtype
        logits = logits.float()

    gates = F.softmax(logits, dim=1)
    metadata["entropy_gating"] = entropy(probs=gates).mean().detach()

    # gates has shape of SE
    num_tokens = gates.shape[0]
    num_experts = gates.shape[1]
    if moe_eval_capacity_token_fraction > 0.0 and eval_mode:
        capacity = math.ceil(moe_eval_capacity_token_fraction * num_tokens)
    else:
        # capacity = capacity_factor * S/E
        capacity = int(capacity_factor * math.ceil(num_tokens / num_experts))

    # Create a mask for 1st's expert per token
    indices1_s = torch.argmax(gates, dim=1)
    mask1 = one_hot(indices1_s, num_classes=num_experts, unsqueeze_indices=True)
    if input_mask is not None and input_mask.any():
        nonpadding = ~input_mask
        mask1 = mask1 * nonpadding.unsqueeze(-1).to(mask1.dtype)

    # for logging (percent of tokens routed to each expert)
    expert1_hist = 100 * torch.histc((indices1_s.squeeze() + 1), bins=num_experts, min=1, max=num_experts) / num_tokens
    metadata["unused_expert1_count"] = (expert1_hist == 0).sum()
    expert1_hist = torch.sort(expert1_hist, dim=0, descending=True).values + torch.finfo(torch.float32).tiny

    sample_count = max(math.ceil(num_experts * SAMPLE_FRACTION), 1)
    metadata["expert1_balance_top"] = expert1_hist[:sample_count].sum()
    metadata["expert1_balance_bottom"] = expert1_hist[-sample_count:].sum()

    gates1_s = (gates * mask1).sum(dim=1)

    # Compute locations in capacity buffer
    locations1 = fused_cumsum_sub_one(mask1)

    # Compute l_aux
    me = torch.mean(gates, dim=0)
    ce = torch.mean(mask1.to(gates.dtype), dim=0)

    l_aux = torch.mean(me * ce)
    l_aux = l_aux * num_experts * num_experts

    if has_tutel:
        locations1_s = torch.sum(locations1 * mask1, dim=1)
        return (
            l_aux,
            metadata,
            capacity,
            num_experts,
            [
                indices1_s,
            ],
            [
                locations1_s,
            ],
            [
                gates1_s,
            ],
        )

    # Remove locations outside capacity from mask
    mask1 = mask1 * torch.lt(locations1, capacity)
    # Store the capacity location for each token
    locations1_s = torch.sum(locations1 * mask1, dim=1)

    # Calculate combine_weights and dispatch_mask
    gates1 = gates1_s.unsqueeze(-1) * mask1.to(gates1_s.dtype)  # einsum("s,se->se")
    # locations1_sc = num_tokens * capacity
    locations1_sc = one_hot(locations1_s, num_classes=capacity, unsqueeze_indices=True)
    combine1_sec = torch.bmm(
        # einsum("se,sc->sec")
        gates1.unsqueeze(-1),
        locations1_sc.to(gates1.dtype).unsqueeze(1),
    )
    dispatch_mask = combine1_sec.bool()
    if use_fp32:
        return l_aux, combine1_sec.to(orig_dtype), dispatch_mask, metadata
    else:
        return l_aux, combine1_sec, dispatch_mask, metadata


class Top1Gate(torch.nn.Module):
    """Gate module which implements Top2Gating as described in Gshard_.
    ::

        gate = Top2Gate(model_dim, num_experts)
        l_aux, combine_weights, dispatch_mask = gate(input)

    .. Gshard_: https://arxiv.org/pdf/2006.16668.pdf

    Args:
        model_dim (int):
            size of model embedding dimension
        num_experts (ints):
            number of experts in model
    """

    wg: torch.nn.Linear

    def __init__(
        self,
        model_dim: int,
        num_experts: int,
        use_fp32=False,
        input_noise_type=None,
        capacity_factor=1.0,
        moe_eval_capacity_token_fraction=EVAL_CAPACITY_TOKEN_FRACTION,
        use_xmoe=False,
    ) -> None:
        # TODO: merge this to top2gate.py
        #
        super().__init__()

        if not use_xmoe:
            self.wg = torch.nn.Linear(model_dim, num_experts, bias=False)
        else:
            self.wg_reduction = torch.nn.Linear(model_dim, 16, bias=False)
            wg = torch.empty(num_experts, 16)
            torch.nn.init.orthogonal_(wg, gain=0.32)
            self.register_parameter("wg", torch.nn.Parameter(wg))

        self.use_xmoe = use_xmoe
        self.use_fp32 = use_fp32
        self.input_noise_type = input_noise_type
        self.capacity_factor = capacity_factor
        self.moe_eval_capacity_token_fraction = moe_eval_capacity_token_fraction

    def forward(self, input, mask=None):  # type: ignore
        if self.use_xmoe:
            input = self.wg_reduction(input)
            with torch.no_grad():
                wg_norm = self.wg.norm(p=2.0, dim=1, keepdim=True)
                self.wg.mul_(1.5 / wg_norm)
            logits = self._cosine(input, self.wg)
            logits = self._make_finite(logits)
        else:
            logits = self.wg(input)

        return top1gating(
            logits,
            mask,
            use_fp32=self.use_fp32,
            capacity_factor=self.capacity_factor,
            eval_mode=not self.training,
            moe_eval_capacity_token_fraction=self.moe_eval_capacity_token_fraction,
            use_xmoe=self.use_xmoe,
            gate_obj=self,
        )

    def _make_finite(self, scores):
        ok = scores.isfinite()
        if not ok.all():
            # NaNs here can break the assignment algorithm
            scores[~ok] = scores[ok].min()
        return scores

    def _get_gating_temperature(self, eps=1e-4):
        if self.gating_t.data.item() < eps:
            return eps
        return self.gating_t

    def _cosine(self, mat1, mat2, eps=1e-4):
        assert mat1.dim() == 2
        assert mat2.dim() == 2
        # mat1 = F.normalize(mat1, p=2.0, dim=1, eps=eps)
        mat2 = F.normalize(mat2.float(), p=2.0, dim=1, eps=eps)
        return mat1.float().matmul(mat2.transpose(0, 1)).type_as(mat1)


gumbel_map: Dict[torch.device, Callable] = {}


def gumbel_rsample(shape: Tuple, device: torch.device) -> Tensor:
    gumbel = gumbel_map.get(device)
    if gumbel is None:
        one = torch.tensor(1.0, device=device)
        zero = torch.tensor(0.0, device=device)
        gumbel = torch.distributions.gumbel.Gumbel(zero, one).rsample  # type: ignore
        gumbel_map[device] = gumbel
    return gumbel(shape)


def one_hot(indices: torch.Tensor, num_classes: int, unsqueeze_indices=False) -> Tensor:
    if unsqueeze_indices:
        indices = indices.unsqueeze(-1)
    assert indices.shape[-1] == 1, "last dimension of indices must be have size 1"
    output = torch.zeros(indices.shape[:-1] + (num_classes,), device=indices.device, dtype=indices.dtype)
    output.scatter_(len(output.shape) - 1, indices, 1)
    return output


def entropy(probs):
    logits = torch.distributions.utils.probs_to_logits(probs)
    p_log_p = probs * logits
    return -p_log_p.sum(-1)


def top2gating(
    logits: torch.Tensor,
    input_mask: Optional[torch.Tensor] = None,
    use_fp32=False,
    second_expert_policy="sampling",
    normalize_gate_prob_before_dropping=False,
    eval_mode=False,
    moe_eval_capacity_token_fraction=0.25,
    batch_prioritized_routing=False,
) -> Tuple[Tensor, Tensor, Tensor]:
    """Implements Top2Gating on logits."""
    metadata = {}
    if use_fp32:
        orig_dtype = logits.dtype
        logits = logits.float()
    gates = F.softmax(logits, dim=1)
    metadata["entropy_gating"] = entropy(probs=gates).mean().detach()
    # gates has shape of SE
    num_tokens = gates.shape[0]
    num_experts = gates.shape[1]
    if moe_eval_capacity_token_fraction > 0.0 and eval_mode:
        capacity = math.ceil(moe_eval_capacity_token_fraction * num_tokens)
    else:
        # capacity = 2S/E
        capacity = 2 * math.ceil(num_tokens / num_experts)

    # Create a mask for 1st's expert per token
    indices1_s = torch.argmax(gates, dim=1, keepdim=True)
    mask1 = one_hot(indices1_s, num_experts)
    if second_expert_policy == "sampling":
        # Create a mask for 2nd's expert per token using Gumbel-max trick
        # https://timvieira.github.io/blog/post/2014/07/31/gumbel-max-trick/
        logits_w_noise = logits + gumbel_rsample(logits.shape, device=logits.device)
    else:
        logits_w_noise = logits
    # Replace top-expert with min value
    logits_except1 = logits_w_noise.masked_fill(mask1.bool(), float("-inf"))
    indices2_s = torch.argmax(logits_except1, dim=1, keepdim=True)
    mask2 = one_hot(indices2_s, num_experts)
    gates1_s = (gates * mask1).sum(dim=1)
    gates2_s = (gates * mask2).sum(dim=1)

    if normalize_gate_prob_before_dropping:
        # Normalize gate probabilities
        denom_s = gates1_s + gates2_s
        # Avoid divide-by-zero
        denom_s = torch.clamp(denom_s, min=torch.finfo(denom_s.dtype).eps)
        gates1_s = gates1_s / denom_s
        gates2_s = gates2_s / denom_s

    if second_expert_policy == "random":
        sampled = (2 * gates2_s) > torch.rand_like(gates2_s)
        mask2 = mask2 * sampled.repeat(num_experts, 1).transpose(1, 0)

    # Compute locations in capacity buffer
    if input_mask is not None and input_mask.any():
        nonpadding = ~input_mask
        mask1 = mask1 * nonpadding.unsqueeze(-1).to(mask1.dtype)
        mask2 = mask2 * nonpadding.unsqueeze(-1).to(mask1.dtype)

    if batch_prioritized_routing:
        # if batch_prioritized_routing:
        importance_scores = -1 * gates.max(dim=1)[0]
        sorted_mask1 = mask1[importance_scores.argsort(dim=0)]
        sorted_cumsum1 = fused_cumsum_sub_one(sorted_mask1) * sorted_mask1
        importance_sorted_locations1 = sorted_cumsum1[importance_scores.argsort(dim=0).argsort(dim=0)]

        sorted_mask2 = mask2[importance_scores.argsort(dim=0)]
        sorted_cumsum2 = fused_cumsum_sub_one(sorted_mask2) * sorted_mask2
        importance_sorted_locations2 = sorted_cumsum2[importance_scores.argsort(dim=0).argsort(dim=0)]

        importance_sorted_locations2 += torch.sum(mask1, dim=0, keepdim=True)

        locations1, locations2 = (
            importance_sorted_locations1,
            importance_sorted_locations2,
        )
    else:
        locations1 = fused_cumsum_sub_one(mask1)
        locations2 = fused_cumsum_sub_one(mask2)
        # Update 2nd's location by accounting for locations of 1st
        locations2 += torch.sum(mask1, dim=0, keepdim=True)

    # Compute l_aux
    me = torch.mean(gates, dim=0)
    ce = torch.mean(mask1.to(gates.dtype), dim=0)
    l_aux = torch.mean(me * ce)
    l_aux = l_aux * num_experts * num_experts

    # for logging purposes
    metadata["overflow_expert1"] = 100 * torch.sum(mask1 * torch.ge(locations1, capacity)) / torch.sum(mask1)
    metadata["overflow_expert2"] = 100 * torch.sum(mask2 * torch.ge(locations2, capacity)) / torch.sum(mask2)

    # Remove locations outside capacity from mask
    mask1_, mask2_ = mask1, mask2
    mask1 = mask1 * torch.lt(locations1, capacity)
    mask2 = mask2 * torch.lt(locations2, capacity)

    # for logging (percent of tokens routed to each expert)
    expert1_hist = 100 * torch.histc((indices1_s.squeeze() + 1), bins=num_experts, min=1, max=num_experts) / num_tokens
    metadata["unused_expert1_count"] = (expert1_hist == 0).sum()
    expert1_hist = torch.sort(expert1_hist, dim=0, descending=True).values + torch.finfo(torch.float32).tiny

    expert2_hist = 100 * torch.histc((indices2_s.squeeze() + 1), bins=num_experts, min=1, max=num_experts) / num_tokens
    metadata["unused_expert2_count"] = (expert2_hist == 0).sum()
    expert2_hist = torch.sort(expert2_hist, dim=0, descending=True).values + torch.finfo(torch.float32).tiny

    sample_count = max(math.ceil(num_experts * SAMPLE_FRACTION), 1)
    metadata["expert1_balance_top"] = expert1_hist[:sample_count].sum()
    metadata["expert1_balance_bottom"] = expert1_hist[-sample_count:].sum()

    metadata["expert2_balance_top"] = expert2_hist[:sample_count].sum()
    metadata["expert2_balance_bottom"] = expert2_hist[-sample_count:].sum()

    if not normalize_gate_prob_before_dropping:
        # Normalize gate probabilities
        gates1_s = (gates * mask1).sum(dim=1)
        gates2_s = (gates * mask2).sum(dim=1)
        denom_s = gates1_s + gates2_s
        # Avoid divide-by-zero
        denom_s = torch.clamp(denom_s, min=torch.finfo(denom_s.dtype).eps)
        gates1_s /= denom_s
        gates2_s /= denom_s

    if has_tutel:
        locations1_s = torch.sum(locations1 * mask1_, dim=1)
        locations2_s = torch.sum(locations2 * mask2_, dim=1)
        return (
            l_aux,
            metadata,
            capacity,
            num_experts,
            [indices1_s, indices2_s],
            [locations1_s, locations2_s],
            [gates1_s, gates2_s],
        )

    # Store the capacity location for each token
    locations1_s = torch.sum(locations1 * mask1, dim=1)
    locations2_s = torch.sum(locations2 * mask2, dim=1)

    # Calculate combine_weights and dispatch_mask
    gates1 = gates1_s.unsqueeze(-1) * mask1.to(gates1_s.dtype)  # einsum("s,se->se")
    gates2 = gates2_s.unsqueeze(-1) * mask2.to(gates2_s.dtype)  # einsum("s,se->se")
    locations1_sc = one_hot(locations1_s, num_classes=capacity, unsqueeze_indices=True)
    locations2_sc = one_hot(locations2_s, num_classes=capacity, unsqueeze_indices=True)
    combine1_sec = torch.bmm(
        # einsum("se,sc->sec")
        gates1.unsqueeze(-1),
        locations1_sc.to(gates1.dtype).unsqueeze(1),
    )
    combine2_sec = torch.bmm(
        # einsum("se,sc->sec")
        gates2.unsqueeze(-1),
        locations2_sc.to(gates2.dtype).unsqueeze(1),
    )
    combine_weights = combine1_sec + combine2_sec
    dispatch_mask = combine_weights.bool()
    if use_fp32:
        return l_aux, combine_weights.to(orig_dtype), dispatch_mask, metadata
    else:
        return l_aux, combine_weights, dispatch_mask, metadata


class Top2Gate(torch.nn.Module):
    """Gate module which implements Top2Gating as described in Gshard_.
    ::

        gate = Top2Gate(model_dim, num_experts)
        l_aux, combine_weights, dispatch_mask = gate(input)

    .. Gshard_: https://arxiv.org/pdf/2006.16668.pdf

    Args:
        model_dim (int):
            size of model embedding dimension
        num_experts (ints):
            number of experts in model
    """

    wg: torch.nn.Linear

    def __init__(
        self,
        model_dim: int,
        num_experts: int,
        use_fp32=False,
        second_expert_policy="sampling",
        normalize_gate_prob_before_dropping=False,
        moe_eval_capacity_token_fraction=0.25,
        batch_prioritized_routing=False,
        use_xmoe=False,
    ) -> None:
        super().__init__()
        if not use_xmoe:
            self.wg = torch.nn.Linear(model_dim, num_experts, bias=False)
        else:
            self.wg_reduction = torch.nn.Linear(model_dim, 16, bias=False)
            wg = torch.empty(num_experts, 16)
            torch.nn.init.orthogonal_(wg, gain=0.32)
            self.register_parameter("wg", torch.nn.Parameter(wg))
        self.use_fp32 = use_fp32
        self.second_expert_policy = second_expert_policy
        self.normalize_gate_prob_before_dropping = normalize_gate_prob_before_dropping
        self.moe_eval_capacity_token_fraction = moe_eval_capacity_token_fraction
        self.batch_prioritized_routing = batch_prioritized_routing
        self.use_xmoe = use_xmoe

    def forward(self, input, mask=None):  # type: ignore
        if self.use_xmoe:
            input = self.wg_reduction(input)
            with torch.no_grad():
                wg_norm = self.wg.norm(p=2.0, dim=1, keepdim=True)
                self.wg.mul_(1.5 / wg_norm)
            logits = self._cosine(input, self.wg)
            logits = self._make_finite(logits)
        else:
            logits = self.wg(input)
        return top2gating(
            logits,
            mask,
            use_fp32=self.use_fp32,
            second_expert_policy=self.second_expert_policy,
            normalize_gate_prob_before_dropping=self.normalize_gate_prob_before_dropping,
            eval_mode=not self.training,
            moe_eval_capacity_token_fraction=self.moe_eval_capacity_token_fraction,
            batch_prioritized_routing=self.batch_prioritized_routing,
        )

    def _cosine(self, mat1, mat2, eps=1e-4):
        assert mat1.dim() == 2
        assert mat2.dim() == 2
        # mat1 = F.normalize(mat1, p=2.0, dim=1, eps=eps)
        mat2 = F.normalize(mat2.float(), p=2.0, dim=1, eps=eps)
        return mat1.float().matmul(mat2.transpose(0, 1)).type_as(mat1)

    def _make_finite(self, scores):
        ok = scores.isfinite()
        if not ok.all():
            # NaNs here can break the assignment algorithm
            scores[~ok] = scores[ok].min()
        return scores